Speaker
Description
Thermographic monitoring is increasingly used to supervise welding, yet robust in-process quality assessment remains challenging because thermal signatures are highly dynamic and depend on interacting physical factors such as heat input, surface condition, emissivity variation, reflections, shielding, and viewing geometry. This paper presents an anomaly-detection workflow for infrared thermography (IRT) sequences recorded during TIG (Tungsten Inert Gas) welding. The goal is to detect atypical thermal behaviour early enough to support operator feedback, reduce scrap, and complement conventional post-process inspection.
A controlled experimental campaign was conducted in which a radiometric infrared camera observed the weld pool and adjacent heat-affected zone during bead-on-plate welding under diverse but representative operating conditions. Multiple materials and parameter settings were deliberately explored to capture natural variability of nominal runs as well as the types of transient disturbances that can appear during arc welding. Each thermographic recording was linked to the corresponding specimen and process metadata to ensure traceability between infrared data and the physical outcome.
The proposed pipeline begins with conversion of the radiometric stream into calibrated temperature maps and selection of a region of interest around the arc–pool area to suppress background effects. Normalization and temporal alignment are applied to make learning less sensitive to changes in absolute temperature level and to highlight structural thermal patterns. Alongside deep learning, the method computes interpretable indicators that serve as baselines and diagnostics, including the evolution of average temperature within the region of interest, inter-frame temperature changes, and the relative extent of high-temperature zones obtained with adaptive thresholding. These descriptors help characterize typical behaviour, identify sudden perturbations, and support curation of nominal data for model training.
For anomalies detection, a convolutional autoencoder is trained exclusively on thermograms representing stable, acceptable process behaviour. The network learns a compact representation of expected spatial temperature distributions and reconstructs incoming frames. During inference, reconstruction error provides an anomaly score: frames that cannot be well reconstructed are treated as potentially abnormal. To reduce false alarms caused by benign variability or optical artifacts, a lightweight post-verification step examines additional image cues (for example, geometry- and edge-related features) to confirm the deviation and provide a coarse qualitative categorization of the event.
Results on the recorded sequences show that peaks in reconstruction error align with visible thermal irregularities and indications of unstable bead formation. By combining physics-informed preprocessing, transparent thermal indicators, and unsupervised deep learning, the workflow offers a practical route toward scalable IR-based supervision of TIG welding in settings where labelled defect data are limited.
